What does First look on start do in the shared code context project, and when should I turn it off?

First look on start in the Tares shared code context project is on by default. It runs the agent once per source repository over the last seven days of commits, so the context repository starts current instead of waiting for the next commit. Turn it off if your context pages are already current or if you want to save model spend.

How it works

First look on start is an option in step 4 of the Shared code context setup wizard, next to the project name, the trigger preview, the model and the round budget. It is on by default.

When you click Start, Tares creates the project’s objects: one commits source per repository, the trigger over them, GitHub’s hosted MCP server registered with your credential, and the agent. With First look on, Tares also runs the agent once for each source repository. Each run covers the last seven days of commits in that repository.

This differs from the normal flow. Without First look, the agent wakes only when a new commit lands on a followed branch. The trigger batches commits per repository, once per five minutes. First look covers the week that came before you started, so the context repository does not stay behind until each repository gets its next commit.

The number of first look runs equals the number of source repositories you ticked. One project takes up to 50 source repositories, so a large project starts with up to 50 runs. Each one is a real agent run that uses your model provider, such as the Anthropic API key you set under Settings > Model providers. The round budget you choose in the same step (12 by default) applies to the agent’s runs.

To see them, open the project page and go to Setup → Advanced setup. Activity lists every run with its status and result, and runs marked first look are the ones from Start. The agent under Agents lists its runs with rounds. On the Overview, What the agents found shows what the agent concluded, one card per change.

Set it up

1. Run Tares locally. See the Quickstart if you have not installed it yet.

tares up

2. Set an Anthropic API key under Settings > Model providers or as ANTHROPIC_API_KEY, and store a GitHub fine-grained token under Settings > GitHub. The token needs Contents and Pull requests at Read and write for the source repositories and the context repository.
3. Open Projects, pick Shared code context and click Set up. Work through the GitHub access, source repositories and context repository steps. The shared code context guide covers each one.
4. In step 4, name the project, pick the model and the round budget, and decide on First look on start.
   • Leave it on if the context pages are stale or empty, or if you want a fresh first pass over the past week.
   • Turn it off if the pages are already current, or if you want to hold back model spend until the first real commit arrives.
5. Check the preview of what Start creates, then click Start. You land on the project page.
6. With First look on, open Setup → Advanced setup and then Activity to follow the runs marked first look. With it off, push a commit to a source repository. Within a minute the source ingests it and the trigger fires for that repository.

What you get

• A context repository that starts current. The agent reads the past seven days of commits per repository at the start, so the pages reflect recent changes from the first day.
• Visible start-up runs. Runs from Start are marked first look in Activity, so you can tell them apart from runs triggered by new commits.
• Control over spend. The option is one switch in the wizard. Leave it on to pay for the initial pass, or turn it off to pay only when commits land.
• The same review flow as later runs. The project’s objects are ordinary objects with their own pages. You can read the agent’s runs and findings the same way for first look and for commit-triggered runs. For the wider picture of how Tares watches systems and escalates, see AI SRE.

When something else fits

• Turn First look off when your context pages already cover the last week of changes, since the agent would mostly conclude that no update is needed.
• Turn it off when you follow many repositories and want to limit model spend, because every source repository means one run.
• Leave it on for a new or neglected context repository, where a first pass over recent commits is the quickest way to catch up.

Related